An analysis of energy-optimized lattice-Boltzmann CFD simulations from the chip to the highly parallel level

Algorithms with low computational intensity show interesting performance and power consumption behavior on multicore processors. We choose the lattice-Boltzmann method (LBM) as a prototype for this scenario in order to show if and how single-chip performance and power characteristics can be generalized to the highly parallel case. LBM is an algorithm for CFD simulations that has gained popularity due to its ease of implementation and suitability for complex geometries. In this paper we perform a thorough analysis of a sparse-lattice LBM implementation on the Intel Sandy Bridge processor. Starting from a single-core performance model we can describe the intra-chip saturation characteristics of the code and its optimal operating point in terms of energy to solution as a function of the propagation method, the clock frequency, and the SIMD vectorization. We then show how these findings may be extrapolated to the massively parallel level on a petascale-class machine, and quantify the energy-saving potential of various optimizations. We find that high single-core performance and a correct choice of the number of cores used on the chip are the essential factors for lowest energy to solution with minimal loss of performance. In the highly parallel case, these guidelines are found to be even more important for fixing the optimal performance-energy operating point, especially when taking the system's baseline power consumption and the MPI communication characteristics into account. Simplistic measures often applied by users and computing centers, such as setting a low clock speed for memory-bound applications, have limited impact.